Re: [BEHAVE] Fwd: IPv6 hosts sending <1280 byte packets

[Iljitsch van Beijnum <[e-mail]>]

On 9 feb 2010, at 21:14, Dan Wing wrote:

>> You can send large packets with IPv4, too.

> Yes, but experience on the IPv4 Internet shows that ICMP packet-too-big are
> unreliable.  So many have given up on IPv4.

I don't know about the latter, but please don't think that issues you see with too bigs in one direction translate to the same issues with too bigs in the other direction.

The reason PMTUD black holes are common is because people filter them out in their firewalls or their hosts because they have some irrational fear of ICMP messages and/or they don't think they need to see those messages. If you have a reduced MTU, you depend on the side with the regular 1500-byte MTU handling PMTUD properly. If they don't, you have a problem that _you_ can't fix. If you use jumboframes and communicate with someone with a smaller MTU, you have to run PMTUD properly _yourself_ to avoid problems. So you would probably not want to filter too big messages. Or if somehow you can't get the PMTUD to work, you can always reduce your MTU to avoid the problem. People with small MTUs generally can't increase those.

So the only issue that I see is the case where too bigs aren't properly generated in the first place. I'm sure there will be a few networks where this is the case, but these are outgoing ICMP messages and as a rule network operators don't filter ICMP, at least not as often as server operators. You also only have a jumbo-clean path towards a limited number of networks, so you don't need 100% of the server operators to handle PMTUD correctly but only a few percent of the network operators. So you will very likely not see any problems. If problems are rare they're also more likely to be fixed when they occur.

Maybe we should ask people on the MTU list about their experiences.